Robust Smith Predictor Design for

Time-Delay Systems with H∞ Performance

Vinicius de Oliveira 1 and Alireza Karimi 2

Automatic Control Laboratory, Ecole Polytechnique Fédérale de

Lausanne (EPFL), Switzerland

Abstract: A new method for robust fixed-order H∞ controller design for uncertain time-delay
systems is presented. It is shown that the H∞ robust performance condition can be represented
by a set of convex constraints with respect to the parameters of a linearly parameterized
primary controller in the Smith predictor structure. Systems with uncertain dead-time and
with multimodel and frequency-domain uncertainty can be considered straightforwardly in the
proposed approach. Furthermore, it is shown that the design method can also be extended to
design of robust gain-scheduled dead-time compensators. The performance of the method is
illustrated by simulation examples.

1. INTRODUCTION uncertainty is proposed by Lee et al. [1999]. In particular,

Most industrial processes present dead time in their dy-
namics. Generally, dead times are caused by the time
needed to transport energy, mass or information, but they
also can be caused by processing time or by accumulation
of time lags in a sequence of simple dynamic systems inter-
connected in series (Normey-Rico and Camacho [2007]).
The presence of dead times in the control loops has two
main consequences: it greatly complicates the analysis and
the design of feedback controllers and it makes satisfac-
tory control performance more difficult to achieve (Palmor
[1996]). Dead-time compensators can be used to improve
the closed-loop performance of classical controllers (PI
or PID controllers) for processes with delay. The Smith
Predictor (SP) (See Fig.1), proposed in the late 1950s
by Smith [1957], was the first dead-time compensation
structure used to improve the performance of the classical
controllers and became the most known and used algo-
rithm to compensate dead time in the industry.
Although the SP offers potential improvement of the
closed-loop performance of process with large dead-time,
it requires a good model since small modeling errors can
lead to very poor performance. In fact, even if the SP is
nominally stabilizing, the closed-loop system may become
unstable for infinitesimal change in the process dynamics
(Palmor [1980]). For this reason, research efforts have been
focused on robustness issues of the SP. A tuning method
for models with one uncertain parameter is proposed by
Brosilow [1979]. Easy tuning rules for SP in the presence
of dead-time uncertainty is addressed in Santacesaria and
Scattolini [1993] and a guideline for selection the closed-
loop bandwidth based on the dead-time uncertainty bound
is proposed. In Palmor and Blau [1994], a robust tuning
rule is developed which considers the modeling error in the
dead time. Robust PID tuning for SP considering model
1 Vinicius de Oliveira is currently a PhD student student at the
Department of Chemical Engineering of the Norwegian University of
Science and Technology, Norway
2 Corresponding author: alireza.karimi@epfl.ch
first and second order plus dead-time systems which may
contain uncertainty in multiple parameters of the model
are considered. In Meinsma and Zwart [2000], tuning
guidelines are presented for setpoint tracking considering
model mismatches in the dead-time.
Recently, many researchers are interested in the optimal
control of dead-time systems, especially H∞ control, i.e.,
to find a controller to internally stabilize the system
and to minimize the H∞ -norm of an associated transfer
function. Many relevant results have been presented in
this framework using modified versions of the SP. See,
for instance, Mirkin [2003], Meinsma and Zwart [2000]
and Zhong [2003a]. Nonetheless, as pointed out by Zhong
[2003b], these controllers are too involved and the predic-
tor always includes additional unstable hidden modes even
for stable plants. As the author argues, these hidden modes
are not safe as they tend to destabilize the system when
implemented. Therefore, the practical significance of these
results are limited.
This paper presents a new method to design fixed-order
SP controllers that considers uncertainty simultaneously in
the dead-time and in the rational part of the model. The
performance specification, like the standard H∞ control
problem, is a constraint on the infinity norm of the
weighted sensitivity function and is represented by a set of
convex constraints in the Nyquist diagram. A line search
on the upper bound γ of the infinity norm of the weighted
sensitivity function can be used in association to a convex
feasibility problem to determine the primary controller
parameters. The proposed method can be used for PID
controllers, which are of great practical interest, as well
as for higher order linearly parametrized controllers in
discrete or continuous time. The main idea is introduced
for LTI-SISO systems and will be extended to MIMO
systems in future works.
It is important to point out that a method to design
fixed-order controllers for standard control loops based on
convex constraints in the Nyquist diagram has been firstly
proposed in Karimi and Galdos [2010]. This paper uses
the same framework to design dead-time compensators for r + u y
time-delay systems. The extension to MIMO systems will Ceq (s) P (s)
−
be based on the idea presented in Galdos et al. [2010b] for
designing H∞ decoupling MIMO controllers.
This paper is organized as follows: In Section II the class of
models, controllers and the control objectives are defined. Fig. 2. Equivalent control diagram
Section III introduces the control design methodology 2.3 Design specifications
based on convex constraints in the Nyquist diagram. In
Section IV the results are extended to unstable time- The SP can be represented by a classical feedback loop
delay systems. Gain-scheduled SP is designed for time- (see Fig. 2) where Ceq (s) is the equivalent controller given
delay systems in Section V. Finally the concluding remarks by
are given. Ceq (s) = C(s)[1 + C(s)H(s)]−1 (6)
with H(s) = Gn (s) − P0 (s) = Gn (s)(1 − e−τn s ).
2. PROBLEM FORMULATION
Let the sensitivity function of the closed-loop system with
2.1 Class of models the i-th plant model Pi (s) be defined as
1 1 + C(s)H(s)
Si (s) = = (7)
Consider the class of stable time-delay LTI-SISO systems 1 + Pi (s)Ceq (s) 1 + C(s)[H(s) + Pi (s)]
with bounded infinity norm. It is assumed that the plant and the complementary sensitivity function as
model can be represented by: Pi (s)Ceq (s) C(s)Pi (s)
P (s) = G(s)e−τ s (1) Ti (s) = = (8)
1 + Pi (s)Ceq (s) 1 + C(s)[H(s) + Pi (s)]
where the time delay τ is unknown but belongs to a finite A standard robust control problem is to design a controller
set {τ1 , τ2 , . . . , τq } and the dead-time free part of the model that satisfies
W1 Si
∞ < 1 for a set of models where
has unstructured multiplicative uncertainty described as: W1 (s) is the performance weighting filter. If the model is
G(s) = Gn (s)[1 + ∆(s)W2 (s)] (2) described by unstructured multiplicative uncertainty, the
where W2 (s) is a known stable uncertainty filter, Gn (s) necessary and sufficient condition for robust performance
the nominal dead-time free model and ∆(s) an unknown is given by (Doyle et al. [1992]):
stable transfer function with
∆
∞ < 1. Therefore, we can
|W1 Si | + |W2 Ti |
∞ < 1 for i = 1, . . . , q (9)
assume that P (s) belongs to a set P of q models given by: The goal of the proposed approach is to design the primary

P = {Pi (s) = G(s)e−τi s ; i = 1, . . . , q} (3) controller C(s) in the SP structure to guarantee robust
performance of the closed-loop system.
2.2 Class of controllers 3. PROPOSED METHOD
The SP control structure shown in Fig.1 is considered. The robust performance condition (9) can be written as:
The nominal model P0 (s) = Gn (s)e−τn s with a fixed τn
belongs to {τ1 , τ2 , . . . , τq } is used for the implementation |W1 (jω)Si (jω)| + |W2 (jω)Ti (jω)| < 1, ∀ω (10)
of the controller. for i = 1, . . . , q. The dependency on frequency ω and s
will be omitted for brevity but the dependency on the
r + u y controller parameter ρ will be highlighted.
C(s) P (s)
−
Let Li (ρ) be defined as Li (ρ) = C(ρ)(H + Pi ). Note that
+ this transfer function is not equal to the open loop transfer
−
Gn (s) e−τn s function Leqi (ρ) = Pi Ceq (ρ) of the equivalent control loop
+
in Fig. 2. The main properties of Li (ρ) are:
yp
+ (1) linearity with respect to the controller parameter ρ;
(2) the closed loop poles are the roots of 1 + Li (ρ) = 0.
Fig. 1. Smith Predictor
The first property is clear because C(ρ) is linearly param-
eterized. To show the second property, let C(ρ) = D Nc
c
,
The primary controller C(s) is linearly parametrized by N
Gn = Dgg be defined. The equivalent controller can be
C(s) = ρ φ(s)
T
(4) written as
where ρ = [ρ1 , ρ2 , . . . , ρnc ] is an nc dimensional vector of
T Nc Dg
Ceq (ρ) = (11)
the controller parameters and Dc Dg + Nc Ng (1 − e−τn s )
φT (s) = [φ1 (s), φ2 (s), . . . , φnc (s)] (5) Then, we have:
is a vector of basis functions with φi (s) transfer functions Dc Dg + Nc Ng (1 − e−τns + e−τi s )
with no RHP poles. For instance, a PID controller could 1 + Leqi (s) = (12)
Dc Dg + Nc Ng (1 − e−τn s )
be linearly parametrized by
Note that the numerator of 1 + Leqi (s) is a quasi-
1 s
ρT = [Kp , Ki , Kd ] , φT (s) = [1, , ] polynomial and its zeros are the closed-loop poles of the
s 1 + Tf s system. On the other hand, we have
 Dc Dg + Nc Ng (1 − e−τns + e−τi s )
1 + Li (ρ) = (13)
Dc Dg
It is clear that 1 + Li (ρ) shares the same zeros with 1 +
Leqi (s). This property will be used to prove the closed loop
stability.
3.1 Main Result
The main result of this section is given in the following
theorem.
Theorem 1. Consider the set of models P in (3) with
multiplicative uncertainty filter W2 (jω), then the linearly
parametrized controller in (4) in the SP structure guaran-
tees closed-loop stability and satisfy the following robust
performance condition:

|W1 Si | + |W2 Ti |
∞ < 1 for i = 1, . . . , q (14)
if

 
W1 (jω)[1 + C(jω, ρ)H(jω)] +
 
W2 (jω)C(jω, ρ)Pi (jω) |1 + Ld (jω)|
− Re{[1 + L∗d (jω)][1 + Li (jω, ρ)]} < 0
∀ω for i = 1, . . . , q (15)
where Ld (jω) is a strictly proper transfer function which
does not encircle the critical point and L∗d (jω) is its
complex conjugate.
Proof : Since the real part of a complex number is less than
or equal to its magnitude, we have
Re{[1 + L∗d ][1 + Li (ρ)]} ≤ |[1 + L∗d ][1 + Li (ρ)]| (16)
Then, using (15) and the fact that |1 + Ld| = |1 + L∗d|, one
obtains
   
W1 (1 + C(ρ)H) + W2 C(ρ)Pi  − |1 + Li (ρ)| < 0
∀ω for i = 1, . . . , q (17)
Using Li (ρ) = C(ρ)(H + Pi ) we have
|W1 (1 + C(ρ)H)| + |W2 C(ρ)Pi |
<1
|1 + C(ρ)(H + Pi )|
∀ω for i = 1, . . . , q (18)
that leads directly to (14). To prove that all closed-loop
transfer functions are stable, consider (15) which gives:
Re{[1 + L∗d (jω)][1 + Li (jω, ρ)]} > 0 ∀ω (19)
or, alternatively,
wno{[1 + L∗d (jω)][1 + Li (jω, ρ)]} = 0 (20)
where wno stands for winding number around the origin.
Since both L∗d (jω) and Li (jω, ρ) are constant or zero
for the semi-circle with infinity radius of the Nyquist
contour the wno depends only on the variation of s in
the imaginary axis. Thus,
wno{[1 + Ld (jω)]} = wno{[1 + Li (jω, ρ)]} (21)
Since Ld (jω) satisfies the Nyquist stability criterion
Li (jω, ρ) will do so and all zeros of 1 + Li (jω, ρ) will be in
the left-hand side of the complex plan. Since the zeros of
1 + Li (jω, ρ) are the closed-loop poles, the system will be
internally stable. 2 A PID primary controller with Tf = 0.01 that minimizes
Remark I: The constraint in (15) is an inner convex
approximation of the non convex constraint in (14) or (10).
The quality of this approximation depends on the choice of
Ld . It can be shown that better approximation is achieved
if Ld is close to Li (ρ) (Karimi and Galdos [2010]). Note
that Li (ρ) can be written as:
Li (ρ) = C(ρ)Gn + C(ρ)(Pi − P0 ) (22)
Therefore, Ld can be seen as the desired open loop transfer
function of the system without dead-time, i.e., C(ρ)Gn .
This will be a good approximation if τi is close to τn or the
model mismatch Pi − P0 is small. For systems with large
uncertainty in the time delay, for each model Pi different
Ld can be considered. In this case, Ldi can be computed
as follows:
Ld
Ldi = Ld + (Pi − P0 ) (23)
Gn
Note that Ldi should not encircle the critical point. An
iterative algorithm can be used to improve the results. In
this algorithm Ld in the (k + 1)-th iteration is computed
using the controller of the last iteration Ck (ρ) and (22):
Ldi = Ck (ρ)Gn + Ck (ρ)(Pi − P0 ) (24)
3.2 Primary controller design
The problem of minimizing the upper bound γ of the infin-
ity norm of the weighted sensitivity function is considered.
Therefore, the primary controller should be obtained from
the following optimization problem:
min γ
ρ
Subject to: (25)

|W1 Si |+|W2 Ti |
∞ < γ for i = 1, . . . , q
This optimization can be convexified using Theorem 1
and solved by an iterative bisection algorithm. At each
iteration j, γj is fixed and W1 and W2 are replaced
by W1 /γj and W2 /γj . Then, a feasibility problem is
solved under the convex constraints (15). If the problem is
feasible, γj+1 is chosen smaller than γj . Otherwise γj+1 is
increased.
Notice that the condition (15) is defined for every fre-
quency ω leading to infinite number of constraints. In
practice, a frequency grid can be used with a sufficiently
large number of frequency points N (a finer grid can
be used around the crossover frequency). The effect of
gridding on the stability and performance of the closed
loop system has been studied in Galdos et al. [2010a].
Example 1 Consider the process described by (1) with
multiplicative uncertainty as in (2) with
1
Gn (s) = (26)
(5s + 1)(10s + 1)
and
−s2 − 2s
W2 (s) = 2 (27)
s + 2s + 1
The unknown time delay τ belongs to the set [4.5, 5, 5.5].
The nominal model used in the SP structure is chosen
as P0 (s) = Gn (s)e−5s . The performance specification is
defined by the following filter:
W1 (s) = 2/(30s + 1)2 (28)

|W1 Si |+|W2 Ti |
∞ < γ for i = 1, 2, 3 should be computed.
Since the controller has an integrator, Ld is chosen as
Ld (s) = ωc /s where ωc = 0.1 rad/s which is 20%
higher than open loop bandwidth. Then, the optimization
problem (25) is solved considering N = 100 equally
spaced frequency points between 10−3 and 103 rad/s. The
resulting primary controller is:
12.3s2 + 3.28s + 0.2201
C(s) = (29)
0.01s2 + s
and leads to γ = 0.313. This controller is compared to
that proposed in Kaya [2001]. Kaya’s controllers performs
better than other controllers presented in the literature
(Palmor and Blau [1994], Hagglund [1992] and Hang et al.
[1995]). Fig. 3 depicts the performance of both controller
on unitary step setpoint change considering the time-delay
τ = 4.5s, τ = 5.0s and τ = 5.5s. As it can be seen, both
controller performed well, however, the proposed controller
achieves faster response.
1
Outputs
C(ρ)
Ceqm (ρ) = (30)
1 + C(ρ)H
where
H = Gm − P0 = (Nm − Nn e−τ s )
1
Dn
Note that the poles of H are zeros of Ceqm (ρ). Therefore,
Nm must be tuned such that the zeros of Nm − Nn e−τ s
cancel the unstable poles in Dn . Once Nm has been
properly designed, the primary controller can be obtained
by solving the optimization problem in (25) redefining
H = Gm − P0 and Li (ρ) = (Gm − P0 + Pi )C(ρ).
Here, care should be taken in the choice of Ld . As it has
been shown, the wno of 1 + Li equals the wno of 1 + Ld .
Therefore, Ld should be chosen such that the number of
encirclement of the critical point (−1 + 0j) by its Nyquist
plot is equal to the number of unstable poles in Pi .

τ = 4.5s
0.5 Example 2 Consider the model studied in Meinsma and
0
Zwart [2000] given by:
k
0 5 10 15 20 25 30 35 40

P (s) = [1 + ∆(s)W2 (s)]e−τ s (31)

1
s−a
τ= 5.0s
where k = 1, a = 1, τ = τn ± 0.02 and τn = 0.2.
Outputs

0.5
The interval of variation of τ is gridded using q = 3
0
0 5 10 15 20 25 30 35 40
equally spaced points. A finer grid just increases the
number of constraints and for this example does not
1
change significantly the final controller. The performance
and uncertainty filters are respectively chosen as:
Outputs

τ= 5.5s
0.5

s+1 s + 1.1
W1 (s) = 2 and W2 (s) = 0.2 (32)
0
0 5 10 15 20
Time [s]
25 30 35 40
10s + 1 s+1
Here, we use the SP with modified dead-time free model
Fig. 3. Example 1: Blue solid line: proposed; black dot- (Fig. 4) due to its simplicity. The dead-time free model
dashed line: ref Kaya [2001] Gm (s) is chosen as
Tm s + 1
Gm (s) = . (33)
s−1
4. EXTENSION TO UNSTABLE SYSTEMS Tm is computed in order to obtain H(s) = Gm (s) − P0 (s)
without a pole in s = 1. Since
The SP in the scheme shown in Fig. 1 cannot be used for 1
unstable plants since the controller will contain zeros in H(s) = [Tm s + 1 − e−0.2s ], (34)
right-hand side of the s-plan which cancel the unstable s−1
poles in the plant and leads to instability. To avoid if Tm = e−0.2 − 1, then s = 1 is a zero of H(s).
this unstable zero-pole cancellation, the control structure A PI as the primary controller is designed. The first step is
shown in Fig.1 should be changed. Several alternatives to choose the transfer function Ld (s), which must encircle
are available in literature to cope with unstable processes the critical point in the Nyquist diagram once and must
with dead-time (see, for example, De Paor [1985], Majhi contain one integrator. Therefore, it is chosen as
and Atherton [1998], Liu et al. [2005], Normey-Rico and s+1
Camacho [2007, 2009]). Consider, for instance, the SP with Ld (s) = 10 . (35)
modified dead-time free model depicted in Fig. 4 which s(s − 1)
is discussed in Normey-Rico and Camacho [2007]. In this Optimization problem (25) is solved considering N = 100
case, the dead-time free model is defined as Gm = N m equally spaced frequency points between ω = 10−3 rad/s
Dn and
the equivalent controller becomes: and ω = 103 rad/s and the following controller is obtained:
C0 (s) = (3.582s + 0.5838)/s (36)
r + u y
C(s) P (s) which yields γ = 0.6854. This result can be further
− improved by using a new Ld (s) based on C0 (s) in the
+ optimization problem. With this new Ld (s) = Gm (s)C0 (s)
Gm (s) P0 (s) − the optimal primary controller is:
C(s) = (2.994s + 0.4612)/s (37)
yp +
and γ = 0.6074. Figure 5 depicts the function
+
Γi (jω) = |W1 (jω)Si (jω)| + |W2 (jω)Ti (jω)|
Fig. 4. Smith Predictor with modified dead-time free where Si and Ti are respectively given by (7) and (8)
model with H = Gm − P0 and Pi is obtained by gridding of τ .
Note that the maximum value of the function is 0.6072, r + u y
which occurs when τ = τn + 0.02 = 0.22, is close to C(s, θ) P (s, θ)
−
the bound γ. It is worth to point out that, although
the conditions given in Theorem 1 are only sufficient to +
−
guarantee
Γi
∞ < γ, with a proper choice of Ld it is Gn (s, θ) e−τn (θ)s
possible to obtain a solution with very low conservatism.
+
Furthermore, the resulting controller is a standard PI yp
which can be implemented in a straightforward manner +
and has great practical significance.
Fig. 6. Gain-Scheduled Smith Predictor
0.65
ρT (θ) = [ρ1 (θ), ρ2 (θ), . . . , ρn (θ)] (40)
0.6
τn+∆τ Every gain is a polynomial function of order δ of the
0.55 scheduling parameters and is defined as
0.5 ρi (θ) = (νi,δ )T θδ + . . . + (νi,1 )T θ + νi,0 (41)
τn−∆τ
Magnitude

0.45 and θ denotes element-by-element power of k of vector θ.

k

0.4 The sensitivity and complementary sensitivity functions

0.35
are respectively given by:
1 + C(s, θ)H(s, θ)
0.3 Si (s, θ) = (42)
0.25
1 + C(s, θ)(H(s, θ) + Pi (s, θ))
C(s, θ)Pi (s, θ)
0.2 Ti (s, θ) = , ∀θ ∈ Θ
1 + C(s, θ)(H(s, θ) + Pi (s, θ))
−2 0 2
10 10 10
Frequency, [rad,s]
where H(s, θ) = Gn (s, θ)−P0 (s, θ). The primary controller
Fig. 5. Example 2: Blue solid line: Γ for τ = 0.18; green is obtained from the following optimization problem:
solid line: Γ for τ = 0.2; black solid line: Blue solid min γ
line: Γ for τ = 0.22; red dot-dashed line: γ. ρ
Subject to: (43)
For the same example, controller designed in Meinsma and
|W1 Si (s, θ)| + |W2 Ti (s, θ)|
∞ < γ
Zwart [2000] leads to the optimal γ = 0.9407 which is for i = 1, . . . , q, ∀θ ∈ Θ
55% higher than the value obtained with the proposed
method. It should be mentioned that the true robust Optimization problem (43) is again solved using an itera-
performance criterion in (25) is not minimized in Meinsma tive bisection algorithm as previously presented. At each
and Zwart [2000]. Instead, the maximum singular value of iteration, a feasibility problem is solved with the following
[W1 (jω)S(jω) W2 (jω)T (jω)] for all ω is minimized by convex constraints:

the H∞ control theory. |W1 (jωk )[1 + C(jωk , θl )H(jωk , θl )|+

5. GAIN-SCHEDULED CONTROLLER DESIGN |W2 (jωk )C(jωk , θl )P (jωk , θl )| |1 + Ld (jωk )|−

Consider an uncertain plant P (s, θ) belonging to the set: Re{[1 + L∗d (jωk )][1 + Li (jωk , θl )]} < 0
for k = 1, . . . , N, i = 1, . . . , q, l = 1, . . . , m (44)
P = {G(s, θ)e−τi (θ)s , i = 1, . . . , q} (38)
where the dead-time free part of the model has unstruc- Example 3 The design method is applied on a simulated
tured multiplicative uncertainty and is described as: system having a resonance whose frequency changes as
G(s, θ) = Gn (s, θ)[1 + ∆(s)W2 (s)] (39) a function of a scheduling parameter θ. Consider the
and θ is a vector of scheduling parameters that belongs following plant model
to a finite set Θ = {θ1 , θ2 , . . . , θm } (corresponding e.g. to P (s, θ) = G(s, θ)e−τ s (45)
the different operating point parameters). It is assumed where G(s, θ) = Gn (s, θ)[1 + ∆(s)W2 (s)] and
that the operating point does not frequently change (the
stability and performance are achieved for the frozen (2 + 0.2θ)2
Gn (s, θ) = (46)
scheduling parameter). The dead-time is also a function s2 + 0.2(2 + 0.2θ)s + (2 + 0.2θ)2
of the scheduling parameter and uncertain, so for a given 1.1337s2 + 6.8857s + 9
value of θ it belongs to the set {τ1 (θ), τ2 (θ), . . . , τq (θ)}. W2 (s) = 0.8 (47)
(s + 1)(s + 10)
We will consider the SP shown in Fig. 6 where both, and θ ∈ [−1, −0.5, 0, 0.5, 1]. Consider also that the dead-
the nominal model P0 (s, θ) = Gn (s, θ)e−τn (θ)s and the time is within the interval τ ∈ [2.7, 3.0, 3.3] but its ex-
primary controller C(s, θ) are functions of the scheduling act value is unknown in runtime. The objective is to
parameter vector θ. The goal is to compute a primary gain- design a primary gain-scheduling PID controller for the
scheduled controller for this scheme that meets the H∞ Smith Predictor structure considering the performance
robust performance specification. 2
filter W1 (s) = (20s+1) 2 . The parameters ρ of the primary

The primary controller C(s, θ) is linearly parametrized by: controller will be affine functions of the scheduling param-
C(s, θ) = ρT (θ)φ(s), where the basis function vector φ(s) eter θ. The filter of the derivative action is chosen to have
is defined in (5) and ρT (θ) is given by a time constant of Tf = 0.01s.
Finally, optimization problem (43) is solved considering
Ld = 1/s and N = 100 equally spaced frequency points
between 10−2 and 102 rad/s. The resulting gain-scheduled
controller is given by: Kp (θ) = −0.0168θ+0.2152, Ki (θ) =
0.0144θ + 2.4736, Kd (θ) = −0.1224θ + 0.6424.
This controller leads to:

|W1 Si (s, θl )| + |W2 Ti (s, θl )|
∞ < γ = 0.8928 (48)
l = 1, . . . , 5, i = 1, 2, 3
The gain-scheduled controller is evaluated considering θ =
−1, 0, 1 and τ = 3.3s. The performance is compared to a
fixed-gain PID designed for the nominal case (θ = 0 and
τ = 3s). Figure 7 shows the step response of the gain-
scheduled controller in all conditions (blue, red and green
solid lines) compared with the fixed PID controller (black
dashed line, highly oscillating).
1 Automatica, 46(8):1388–1394, 2010.
0.9
0.8
0.7
0.6
Outputs
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15
Time [s]
20 25 30
Fig. 7. Example 3: Blue, red and green solid line: gain-
scheduled PID Smith Predictor and G2 using θ1 =
−1, θ3 = 0 and θ5 = 1 respectively; black dashed line:
fixed PID Smith Predictor using θ = −1.
6. CONCLUSIONS
This paper presents a new method to design robust Smith
predictor for uncertain time-delay systems using convex
optimization techniques. The proposed approached allows
one to design PI/PID as well as higher order primary
controllers in the Smith predictor structure which provide
robust H∞ performance for systems with uncertain dead-
time and multiplicative or multimodel uncertainty in the
dead-time free model of the system. The method is based
on a convex approximation of the H∞ robust performance
criterion in the Nyquist diagram. This approximation
relies on the choice of a desired open-loop transfer function
Ld for the dead-time free model of the plant. The extension
of the approach to MIMO systems is under investigation.
REFERENCES
C. B. Brosilow. The structure and design of Smith
predictors from the viewpoint of inferential control.
In Proceedings of Joint American Control Conference,
Denver, Colorado, 1979.
A. M. De Paor. A modified Smith predictor and controlled
for unstable processes with time delay. International
Journal of Control, 41(4):1025–1036, 1985.
C. J. Doyle, B. A. Francis, and A. R. Tannenbaum.
Feedback Control Theory. Mc Millan, New York, 1992.
G. Galdos, A. Karimi, and R. Longchamp. Robust con-
troller design by convex optimization based on finite
frequency samples of spectral models. In 49th IEEE
Conference on Decision and Control, Atlanta, USA,
2010a.
G. Galdos, A. Karimi, and R. Longchamp. H∞ controller
design for spectral MIMO models by convex optimiza-
tion. Journal of Process Control, 20(10):1175 – 1182,
2010b.
T. Hagglund. A predictive PI controller for processes with
long dead times. IEEE Control Systems Magazine, 12
(1):57–60, 1992.
C. C. Hang, Q. Wang, and L. S. Cao. Self-tuning Smith
predictors for processes with long dead time. Interna-
tional journal of adaptive control and signal processing,
9(3):255–270, 1995.
A. Karimi and G. Galdos. Fixed-order H∞ controller de-
sign for nonparametric models by convex optimization.
I. Kaya. Tuning Smith predictors using simple formulas
derived from optimal responses. Industrial & engineer-
ing chemistry research, 40(12):2654–2659, 2001.
D. Lee, M. Lee, S. Sung, and I. Lee. Robust PID tuning
for Smith predictor in the presence of model uncertainty.
Journal of Process Control, 9(1):79–85, 1999.
T. Liu, Y. Z. Cai, D. Y. Gu, and W. D. Zhang. New
modified Smith predictor scheme for integrating and
unstable processes with time delay. IEE Proceedings-
Control Theory and Applications, 152(2):238–246, 2005.
S. Majhi and D. P. Atherton. A new Smith predictor and
controller for unstable and integrating processes with
time delay. In Proceedings of the 37th IEEE Conference
on Decision and Control, pages 1341–1345, 1998.
G. Meinsma and H. Zwart. On H∞ control for dead-time
systems. IEEE Transactions on Automatic Control, 45
(2):272–285, 2000.
L. Mirkin. On the extraction of dead-time controllers
and estimators from delay-free parametrizations. IEEE
Trans. on Automatic Control, 48(4):543–553, 2003.
J. E. Normey-Rico and E. F. Camacho. Control of dead-
time processes. Springer Verlag, 2007.
J. E. Normey-Rico and E. F. Camacho. Unified approach
for robust dead-time compensator design. Journal of
Process Control, 19(1):38–47, 2009.
Z. J. Palmor. Stability properties of Smith dead-time com-
pensator controllers. International Journal of Control,
32:937–49, 1980.
Z. J. Palmor. Time-delay compensation Smith predictor
and its modifications. In W. Levine, editor, The Control
Handbook. CRC Press, Boca Raton, FL, 1996.
Z. J. Palmor and M. Blau. An auto-tuner for Smith dead
time compensator. International Journal of Control, 60
(1):117–135, 1994.
C. Santacesaria and R. Scattolini. Easy tuning of Smith
predictor in presence of delay uncertainty. Automatica,
29(6):1595–1597, 1993.
O. J. M. Smith. Closer control of loops with dead time.
Chemical Engineering Progress, 53(5):217–219, 1957.
Q. C. Zhong. On standard H∞ control of processes with a
single delay. IEEE Transactions on Automatic Control,
48(6):1097–1103, 2003a.
Q. C. Zhong. H∞ control of dead-time systems based on
a transformation. Automatica, 39(2):361–366, 2003b.